Short answer: Dessa is a full-duplex speech-to-speech AI from Block, the company behind Square and Cash App, announced on 23 October 2026. It listens and talks at the same time, was trained only on recordings of human conversation, and is not available to the public yet. The generator above is not Dessa: it is Grok text to speech, which you can use today to make speech with laughs, sighs and whispers. Leave your email at the top of the tool and we send one message when Dessa opens to the public.
Dessa at a glance
| Dessa | |
|---|---|
| Made by | A small AI research lab at Block that reports to Jack Dorsey |
| Announced | 23 October 2026, as a research preview |
| Type | Full-duplex speech to speech: audio in, audio out, listening while it talks |
| Training data | Human conversation audio only; no text, no pretrained models |
| Training compute | One NVIDIA GB200 NVL72 system |
| Runs on | A MacBook Pro with M4 Max or better, locally; faster than real time |
| Access | Block’s own waitlist; early access for AI safety researchers. Get our launch alert |
| API, price, weights | None announced |
| Parameters, languages | Not published |
Checked 26 Oct 2026 against Block's Dessa page and TechCrunch's launch report. We will update this table when Block opens access.
What makes Dessa different
Most voice assistants run a chain: speech to text, a text language model, then text back to speech. Even newer speech-to-speech models usually start from a text model. Block says Dessa skips text entirely and learned language "simply by listening", from what it calls hundreds of times fewer words than a large language model reads.
Two things follow from that, going by Block's demos:
- It behaves like someone on a call. Because it is full-duplex, it can backchannel ("yeah!", "got it", "ooh") while you are still talking, and it learned from real conversations when to cut in.
- Emotion is learned, not tagged. Block contrasts Dessa with models such as ElevenLabs v4 and Fish Audio Drama 3, where you write a tag to get a laugh or a sigh. Dessa picks up sarcasm, laughter, a frustrated sigh or a voice cracking with sadness from the conversation itself.
Block is also open about the risk: a voice this human can mislead people, which is why the first outside users are safety researchers looking for ways it could deceive a listener.
Where the name comes from
The lab was started by former members of Dessa, a Toronto machine-learning startup best known for deepfake-detection research, which Square acquired in February 2020. Square renamed itself Block in 2021, and the model now carries the old company's name.
Dessa vs Grok voice models
| Dessa (Block) | Grok Voice 3.0 (xAI) | Grok TTS (xAI) | |
|---|---|---|---|
| What it does | Live two-way conversation | Live voice agent | Reads text aloud |
| Can you use it today | No, waitlist | Yes, API | Yes, API and free here |
| Price | Not announced | $0.08 per minute | $15 per million characters |
| How emotion is set | Learned from conversation | Voice choice and session settings | Speech tags in the text |
| Where it runs | Locally (M4 Max or better) | xAI cloud | xAI cloud |
| Voices | Not published | 28 voices | 28 voices |
If you need a voice agent that picks up the phone today, see the Grok Voice 3.0 guide and the cost calculator. If you need expressive narration, a script read with feeling, or a quick sound check, text to speech is the tool, and the box above is free to try.
Make Dessa-style expressive speech with tags
Dessa decides on its own where to laugh. With Grok TTS you place each sound yourself, which is less spontaneous but fully predictable. The sample above already uses three tags; press Generate to hear it. Tags that get closest to the moments in Block's demos:
- Laughter:
[laugh],[chuckle],[giggle] - Sighs and breathing:
[sigh],[breath],[exhale] - Sadness:
[cry], or wrap a line in<soft>…</soft>and<slow>…</slow> - Asides:
<whisper>…</whisper>for a line said under the breath - Thinking pauses:
[pause]and[long-pause]between clauses
Write the way people talk: short sentences, "yeah" and "oh" where a person would say them, and a tag only where you would actually hear the sound. The full list, with what each tag does, is in the Grok TTS speech tags guide.
Signed out, you get 3 previews a day of up to 200 characters, played on the page; a free account adds MP3 download, and a Google account raises the limit to 1,000 characters per clip. Your text goes to xAI's model through our Cloudflare function to make the audio and is not stored.
This page is an independent explainer. We are not affiliated with Block, Square, Cash App or xAI, and have no early access to Dessa.
Last updated: